Pith. sign in

REVIEW 5 major objections 4 minor 49 references

SEFD: Semantic-Enhanced Framework for Detecting LLM-Generated Text

T0 review · 5 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read SEFD fuses retrieval similarity with existing detectors to catch paraphrased LLM text.

desk verdict SEFD's paraphrase gains are real but conditional on a sequential-ordering assumption; the experiments stack the deck and the paper never tests the failure case. read the letter →

arxiv 2411.12764 v1 pith:MLSLBG7W submitted 2024-11-17 cs.CL cs.AIcs.IR

Weiqing He , Bojian Hou , Tianqi Shang , Davoud Ataee Tarzanagh , Qi Long , Li Shen This is my paper · ORCID
classification cs.CLcs.AIcs.IR
keywords LLM-generatedtextdetectionparaphrasingattackssemanticsimilarityretrievalpoolupdatingsequentialfusionfunctionwatermarkingzero-shotdetectors
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that an existing LLM-generated-text detector can be made substantially more resistant to paraphrase attacks by adding a retrieval step that compares each incoming text, by semantic similarity, against a running pool of previously seen LLM outputs. The resulting wrapper, SEFD (Semantic-Enhanced Framework for Detecting LLM-Generated Text), fuses the detector's score with the best cosine similarity to anything in the pool and updates the pool as texts arrive, so a paraphrase can be matched to its original before the paraphrase is classified. Experiments across four source LLMs and four detection methods report gains in paraphrase detection accuracy while preserving performance on ordinary LLM-generated text. The intended setting is sequential text, such as chronologically ordered forum comments or Q&A answers.

What carries the argument

The load-bearing object is the retrieval pool together with the fusion rule. Each input text $x_i$ is embedded, and its similarity score $s_{\mathrm{sim}}$ is the maximum cosine similarity against embeddings already in the pool. The final score is $s = s_{\mathrm{det}}/(1+10^{-\lambda_1}-s_{\mathrm{sim}})^{1/\lambda_2}$, which leaves the detector score nearly unchanged when similarity is near $0$ and amplifies it by about $10^{\lambda_1/\lambda_2}$ when similarity is near $1$. The updating rule compares detection and similarity scores against two thresholds and decides whether to ignore the text, add it to the pool, or replace the closest pool entry; the replacement case is what lets the pool track recursive paraphrases.

What would settle it

Shuffle a test sequence so that paraphrased answers appear before their originals and keep every other setting identical; if paraphrase detection accuracy does not fall back to the level of the initial detector alone, then the retrieval-pool mechanism described is not the cause of the gain. A second check would be to insert a human-written answer that is semantically close to an LLM output already in the pool and see whether the fusion function misclassifies it.

Watch

Extended reading notes

Core claim

SEFD's central claim is that paraphrasing weakens standard detection signals but leaves semantic content largely intact, and that a retrieval pool of previously seen LLM outputs supplies exactly the signal the detector loses. By combining a detector score with a maximum cosine-similarity score through the fusion function $s_{\mathrm{det}}/(1+10^{-\lambda_1}-s_{\mathrm{sim}})^{1/\lambda_2}$, the framework amplifies weak detector scores precisely when the input is semantically close to a known LLM output. The updating rule keeps the pool current, replacing old entries with their paraphrased versions so recursive paraphrases remain detectable. On the paper's datasets this recovers much of the paraphrase-induced accuracy drop across all four tested detectors and all four source models.

Load-bearing premise

The gain depends on every paraphrased text arriving after its original LLM-generated text, so the retrieval pool has already seen (or can immediately store) a semantically matching anchor before the paraphrase is classified.

Editorial extensions

If this is right

  • Paraphrase detection accuracy improves relative to the initial detector alone across all datasets tested, with the largest relative gains for detectors whose scores drop most under paraphrase.
  • An empty initial pool is sufficient for substantial gains, because the updating rule adds newly seen LLM-generated texts and later matches their paraphrases against them.
  • Larger initial pools generally improve detection accuracy, though a non-robust encoder can occasionally make a non-empty pool hurt human-text classification.
  • Recursive paraphrasing, which pushes baseline detection accuracy close to zero, causes only a moderate decline under SEFD.
  • SEFD is detector-agnostic: any detector with workable thresholds can be wrapped, so the framework does not depend on the four specific detectors tested.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the ordering assumption fails and a paraphrase arrives before its original is seen, or the original never appears in the stream, SEFD should fall back to the initial detector; this is testable by shuffling the input order.
  • The same wrapper logic could be transferred to other meaning-preserving transformations, such as machine translation or summarization, whenever a detector's score degrades but semantic similarity survives.
  • Contamination risk is inherent: a false positive in the initial detector adds a human text to the pool, which could raise later similarity scores for other human texts; low false-positive-rate operating points mitigate but do not remove this risk.
  • The four tuning parameters are detector-specific and the paper does not claim global optimality, so deployment would require a calibration step for each new detector and data distribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper introduces SEFD, a wrapper framework that augments any existing LLM-generated-text detector with a retrieval component. For each input text, SEFD computes an initial detection score s_det, a maximum-cosine semantic-similarity score s_sim against a retrieval pool of embeddings of previously seen LLM-generated texts, and fuses them via Eq. (4) to produce a final score. A four-case updating rule (Table I) adds or replaces entries in the retrieval pool. Experiments on long-form QA data from Reddit, using GPT-2 XL, OPT-13B, GPT-3.5, and GPT-4o-mini as source models and Log-Likelihood, DetectGPT, ID-MLE, and Soft Watermarking as initial detectors, report improved AUROC and detection accuracy for paraphrased text and improved resistance to recursive paraphrasing. The central claim is that SEFD substantially enhances detection accuracy in paraphrasing scenarios while maintaining robustness for standard LLM-generated content.

Significance. If the reported gains hold under realistic streaming conditions, SEFD is a practically useful, detector-agnostic wrapper whose mechanism is easy to understand: a paraphrase of a previously seen LLM output retains high semantic similarity to the stored embedding, so retrieval can compensate for the initial detector's degraded paraphrase score. The paper's strengths are its breadth (four initial detectors, four source LLMs, and a recursive-paraphrase attack scenario), its clear separation of the fusion and updating components, and the public code link. The main caveat is that the headline numbers are produced under deliberately favorable conditions: the input stream is ordered so that every source text precedes its paraphrase, and one of the two main configurations seeds the retrieval pool with one fifth of the test LLM texts. As a result, the experiments demonstrate feasibility under favorable conditions rather than a validated general solution.

major comments (5)
  1. [Section IV-B / III-D] The 'Input data summary' in Section IV-B structures each input sequence as: all LLM-generated texts, then all human-written answers, then all paraphrased versions. This guarantees Assumption A.2 in its strongest form: by the time any paraphrase is scored, the retrieval pool has already had the opportunity to store the exact source text (via Situation 2 of the updating rule when the pool starts empty, or via the initial pool when it is non-empty). The reported paraphrase gains therefore measure SEFD under the maximally favorable ordering. The paper does not test, report, or even state the boundary condition that SEFD's paraphrase benefit requires the source of the paraphrase to have appeared earlier in the same processed stream; when that condition fails, s_sim has no matching anchor and SEFD degenerates to the initial detector, whose paraphrase accuracy is low (e.g., DetectGPT at 8.72 DA for GPT-2 XL in Table III). Please add experiments with shuffled or realistic orderings and state this boundary condition in the limitation section (V-A).
  2. [Section IV-C / Table III] In the 'pool size 1/5' condition, the initial retrieval pool contains 1/5 of the LLM-generated test set (600 of 3000 responses per dataset). For those 600 original texts and for paraphrases derived from them, a high s_sim is obtained simply because the exact or near-identical source embedding is in the pool at decision time. The table does not separate the accuracy contribution from these 'seen' sources from the contribution for unseen sources, so the 'pool size 1/5' rows conflate retrieval from the test set with generalization. The same leakage affects the recursive-paraphrase experiment in Section IV-C3, which also uses a 1/5 initial pool. Please evaluate with a disjoint retrieval corpus or, at minimum, report accuracy separately for texts whose source is and is not contained in the initial pool.
  3. [Section III-C / Table II] The four free parameters (epsilon_det, epsilon_sim, lambda1, lambda2) are fixed to single values per detector (Table II), selected on the same data used for evaluation, and Section IV-B states that they 'may not yield the optimal results.' No validation procedure, sensitivity analysis, or run-to-run variance is reported, so it is unclear whether the improvements in Table III are robust to reasonable parameter perturbations. Since the fusion function in Eq. (4) amplifies s_det by approximately 10^{lambda1/lambda2} when s_sim is near 1, the parameter sensitivity is potentially large. Please add a validation protocol, error bars or confidence intervals, and a sensitivity analysis over the thresholds.
  4. [Section III-C / IV-C] The paper never reports the performance of the retrieval-only similarity score (thresholding s_sim alone) as a baseline. Because the entire claimed gain over the initial detector comes from the retrieval component, this baseline is necessary to determine whether the fusion function provides any benefit beyond simply using s_sim, and to contextualize the 'systematic integration' claim. Please add a retrieval-only row to Table III and, if applicable, to the recursive-paraphrase experiment.
  5. [Abstract / IV-C1 / Table III] The abstract's statement that SEFD maintains robustness for standard LLM-generated content is not supported by the empty-pool configuration: in Table III, original-text detection accuracy decreases relative to the initial detector for Log-Likelihood (GPT-2 XL: 49.4 to 38.0), DetectGPT (GPT-2 XL: 58.54 to 54.56), and ID-MLE (GPT-3.5: 4.1 to 3.0). Section IV-C1 acknowledges this in passing, but the unqualified abstract claim should be revised to distinguish the empty-pool and non-empty-pool regimes.
minor comments (4)
  1. [Algorithm 1] Algorithm 1, lines 13-16, has an inconsistency: the second and fourth branches of the updating rule compare si_det against both thresholds (e.g., 'if si_det >= epsilon_det and si_det < epsilon_sim'), but Table I and Section III-D clearly intend si_sim in the threshold comparisons. As printed, the pseudocode does not implement the stated updating rule.
  2. [Section IV-C3] The recursive-paraphrase subsection says 'The result is presented in Fig. ??'; the figure reference is unresolved and should be fixed to the actual figure number.
  3. [Section IV-A] The metric called 'detection accuracy' is actually TPR at 1% FPR; please define this explicitly and consider renaming it to avoid confusion with standard classification accuracy, since the latter would be a different quantity.
  4. [Section III-B] The use of maximum cosine similarity over a large pool can inflate s_sim for human text when the pool contains many answers to the same question; the paper should justify this max-pooling choice or report additional statistics (e.g., mean or top-k similarity) to show that the separation in Fig. 4 is not an artifact of the max operator.

Circularity Check

2 steps flagged · score 6.0 of 10

Partial circularity in the original-text results: retrieval-pool seeding and self-comparison make part of the reported gains lookup-by-construction, while the paraphrase core remains independently supported.

  1. self definitional [Section III-B, Eq. (3) and Fig. 4]
    "For each LLM, we initialize the retrieval pool that includes all the original answers. As shown in Fig. 4, the similarity scores of LLM-generated text are concentrated around 1."

    Eq. (3) defines s_sim as the maximum cosine similarity between the candidate embedding and all embeddings in the retrieval pool. Since the pool is initialized with all original LLM answers, every original LLM text is compared with itself, forcing its cosine similarity to 1 (or near 1) regardless of the encoder or detector. The 'concentration around 1' in Fig. 4 for original LLM text is therefore a self-comparison artifact of the definition, not an independent empirical demonstration that retrieval generalizes. The paraphrased-text scores in the same figure are not circular, because paraphrases are not stored, but the motivating evidence for the retrieval component is partly tautological.

  2. fitted input called prediction [Section IV-C1, Table III and preceding paragraph]
    "Intuitively, a larger pool size generally leads to better performance since when detecting text already present in the pool, the similarity score approaches 1, allowing for accurate classification."

    The '+ pool size -1/5' condition preloads one fifth of the LLM-generated test texts into the retrieval pool, and the same texts then reappear as test inputs. The reported improvement in 'original' LLM-text detection at this setting is therefore partly exact-match retrieval of seeded instances rather than detection of unseen text; the quoted sentence concedes that the mechanism is 'text already present in the pool.' This makes the original-text component of Table III partly lookup-by-construction, although the paraphrase results, especially with pool size 0 and recursive paraphrasing, retain independent content.

full rationale

SEFD is an empirical wrapper rather than a first-principles derivation, and the central paraphrase claim has independent content: with an empty initial pool, paraphrases are still detected because the updating rule stores originals before the paraphrases arrive, and the recursive-paraphrase experiment tests generalization beyond stored items. However, two places reduce by construction. First, the motivating Fig. 4 stores every original in the pool and then reports that original LLM text has similarity near 1; by Eq. (3), the maximum cosine includes the text compared with itself, so this is a tautology. Second, Table III's '+ pool size -1/5' condition preloads one fifth of the test LLM originals into the pool, and the paper's own explanation ('when detecting text already present in the pool, the similarity score approaches 1') confirms that part of the original-text gain is exact-match lookup rather than generalization. These issues do not make the paraphrase claims circular, but they do mean the 'robustness for standard LLM-generated content' result is partly an artifact of seeding. The framework's real-world applicability is also conditional on Assumption A.2, which the experiments enforce by ordering all LLM originals before all paraphrases; that is a boundary condition rather than a circularity. Author self-citations ([2]-[6]) appear only in the introduction and are not load-bearing.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The framework's central claim rests on several hand-set parameters and domain assumptions about sequential text. No new physical or conceptual entity is introduced. The retrieval pool is a data structure, not an invented entity. The count of free parameters and assumptions is high relative to the simplicity of the method, which limits how strongly the empirical gains can be generalized.

free parameters (5)
  • epsilon_det (detection threshold for pool update) = Per detector in Table II: Log-likelihood -2.5; DetectGPT 0.5; ID-MLE -11; Soft Watermarking 4
    Used in the update rule to decide whether to add or replace a pool entry; chosen by hand per detector, with no validation split described.
  • epsilon_sim (similarity threshold for pool update) = 0.85
    Fixed similarity threshold across all experiments; selected manually to separate LLM-like from human-like similarity scores.
  • lambda1 = 1
    Fusion parameter controlling amplification when similarity score is close to 1; fixed to 1 in all experiments.
  • lambda2 = 6
    Fusion parameter controlling how strongly similarity score scales the detection score; fixed to 6 in all experiments.
  • initial retrieval pool size M0 = Main results use 1/5 of LLM-generated texts; sweep from empty to full in Fig. 6
    Chosen by hand as an experimental variable; performance depends on it, and no principled selection rule is provided.
assumptions (5)
  • domain assumption Inputs arrive strictly sequentially (A.1), one text at a time.
    Section III-D; the pool update requires examining texts in order, so concurrent or batch arrival is outside the method's scope.
  • domain assumption For any LLM-generated text, its paraphrased version always comes after the original (A.2).
    Section III-D; the retrieval defense only works if the original can enter the pool before the paraphrase is scored; this is the load-bearing ordering assumption.
  • domain assumption Human-written and LLM-generated text spaces are disjoint, and any paraphrased text is classified as LLM-generated.
    Section III; simplifies the binary task but ignores the ambiguous case of human text rewritten by an LLM, which the limitations section acknowledges.
  • domain assumption Paraphrasing preserves enough semantic content that cosine similarity remains high.
    Section III-B and III-D; the entire similarity signal depends on this, and the authors note recursive paraphrasing erodes it.
  • domain assumption The sentence encoder (all-MiniLM-L6-v2) produces embeddings where semantic similarity reliably separates LLM answers from human answers.
    Section III-B and IV-C1; the authors acknowledge the encoder can be non-robust and cause false matches for human answers to the same questions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SEFD: Semantic-Enhanced Framework for Detecting LLM-Generated Text." pith.science (2026). https://pith.science/paper/MLSLBG7W

@misc{pith2026241112764,
  author       = {Pith},
  title        = {Pith review of: SEFD: Semantic-Enhanced Framework for Detecting LLM-Generated Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MLSLBG7W}},
  note         = {Machine review of arXiv:2411.12764}
}
read the original abstract

The widespread adoption of large language models (LLMs) has created an urgent need for robust tools to detect LLM-generated text, especially in light of \textit{paraphrasing} techniques that often evade existing detection methods. To address this challenge, we present a novel semantic-enhanced framework for detecting LLM-generated text (SEFD) that leverages a retrieval-based mechanism to fully utilize text semantics. Our framework improves upon existing detection methods by systematically integrating retrieval-based techniques with traditional detectors, employing a carefully curated retrieval mechanism that strikes a balance between comprehensive coverage and computational efficiency. We showcase the effectiveness of our approach in sequential text scenarios common in real-world applications, such as online forums and Q\&A platforms. Through comprehensive experiments across various LLM-generated texts and detection methods, we demonstrate that our framework substantially enhances detection accuracy in paraphrasing scenarios while maintaining robustness for standard LLM-generated content.

Figures

Figures reproduced from arXiv: 2411.12764 by the authors.

Figure 1
Figure 1. A brief version of SEFD structure. SEFD comprises a retrieval [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The detailed structure of our framework. The input sequence on the left consists of three texts: the first is generated by an LLM, the second is [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Detection score distributions for four different detectors: Log [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Similarity score distribution on four different LLM generated datasets: [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Recursive paraphrasing the process is summarized in Alg. 1. There are four situations presented in Table I, for the sake of narrative we define the text in the retrieval pool3 that has the highest similarity score with xi as yτ , then we explain the details as follow: …
Figure 6
Figure 6. Figure 6: Detection accuracy under different initial pool sizes with Detect [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Detection performance of SEFD under recursive paraphrasing attacks. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 23 canonical work pages

  1. [24]

    Para- phrasing evades detectors of ai-generated text, but retrieval is an effective defense,

    K. Krishna, Y . Song, M. Karpinska, J. Wieting, and M. Iyyer, “Para- phrasing evades detectors of ai-generated text, but retrieval is an effective defense,” arXiv preprint arXiv:2303.13408 , 2023

    arXiv 2023

  2. [1]

    Truthfulqa: Measuring how models mimic human falsehoods,

    S. Lin, J. Hilton, and O. Evans, “Truthfulqa: Measuring how models mimic human falsehoods,” arXiv preprint arXiv:2109.07958 , 2021

    arXiv 2021

  3. [2]

    Interpretability study for long interview transcripts from behavior intervention sessions for family caregivers of dementia patients,

    W. He, B. Hou, G. Demiris, and L. Shen, “Interpretability study for long interview transcripts from behavior intervention sessions for family caregivers of dementia patients,” AMIA Summits on Translational Science Proceedings, vol. 2024, p. 201, 2024

    work page 2024

  4. [3]

    Leveraging social determinants of health in alzheimer’s research using llm-augmented literature mining and knowledge graphs,

    T. Shang, S. Yang, W. He, T. Zhai, D. Li, B. Hou, T. Chen, J. H. Moore, M. D. Ritchie, and L. Shen, “Leveraging social determinants of health in alzheimer’s research using llm-augmented literature mining and knowledge graphs,” arXiv preprint arXiv:2410.09080 , 2024

    arXiv

  5. [4]

    Learning with interpretable structure from gated rnn,

    B.-J. Hou and Z.-H. Zhou, “Learning with interpretable structure from gated rnn,” IEEE transactions on neural networks and learning systems , vol. 31, no. 7, pp. 2267–2279, 2020

    work page 2020

  6. [5]

    Learning with feature evolvable streams,

    B.-J. Hou, L. Zhang, and Z.-H. Zhou, “Learning with feature evolvable streams,” Advances in Neural Information Processing Systems , vol. 30, 2017

    work page 2017

  7. [6]

    Interpretable deep clustering survival machines for alzheimer’s disease subtype discovery,

    B. Hou, Z. Wen, J. Bao, R. Zhang, B. Tong, S. Yang, J. Wen, Y . Cui, J. H. Moore, A. J. Saykin, et al., “Interpretable deep clustering survival machines for alzheimer’s disease subtype discovery,” Medical Image Analysis, p. 103231, 2024

    work page 2024

  8. [7]

    A watermark for large language models,

    J. Kirchenbauer, J. Geiping, Y . Wen, J. Katz, I. Miers, and T. Gold- stein, “A watermark for large language models,” arXiv preprint arXiv:2301.10226, 2023

    arXiv 2023

Show all 49 references
  1. [8]

    Adversarial watermarking transformer: Towards tracing text provenance with data hiding,

    S. Abdelnabi and M. Fritz, “Adversarial watermarking transformer: Towards tracing text provenance with data hiding,” in 2021 IEEE Symposium on Security and Privacy (SP) , pp. 121–140, IEEE, 2021

    2021

  2. [9]

    The ethical need for watermarks in machine-generated language,

    A. Grinbaum and L. Adomaitis, “The ethical need for watermarks in machine-generated language,” arXiv preprint arXiv:2209.03118 , 2022

    2022 arXiv

  3. [10]

    Eagle: A domain generalization framework for ai-generated text detection,

    A. Bhattacharjee, R. Moraffah, J. Garland, and H. Liu, “Eagle: A domain generalization framework for ai-generated text detection,” arXiv preprint arXiv:2403.15690, 2024

    2024 arXiv

  4. [11]

    The enemy in your own camp: How well can we detect statistically-generated fake reviews–an adversarial study,

    D. Hovy, “The enemy in your own camp: How well can we detect statistically-generated fake reviews–an adversarial study,” inProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pp. 351–356, 2016

    2016

  5. [12]

    Defending against neural fake news,

    R. Zellers, A. Holtzman, H. Rashkin, Y . Bisk, A. Farhadi, F. Roesner, and Y . Choi, “Defending against neural fake news,” Advances in neural information processing systems , vol. 32, 2019

    2019

  6. [13]

    Real or fake? learning to discriminate machine from human generated text,

    A. Bakhtin, S. Gross, M. Ott, Y . Deng, M. Ranzato, and A. Szlam, “Real or fake? learning to discriminate machine from human generated text,” arXiv preprint arXiv:1906.03351 , 2019

    1906 arXiv

  7. [14]

    Automatic detection of generated text is easiest when humans are fooled,

    D. Ippolito, D. Duckworth, C. Callison-Burch, and D. Eck, “Automatic detection of generated text is easiest when humans are fooled,” arXiv preprint arXiv:1911.00650, 2019

    1911 arXiv

  8. [15]

    New ai classifier for indicating ai-written text,

    OpenAI, “New ai classifier for indicating ai-written text,” 2023

    2023

  9. [16]

    Detectgpt: Zero-shot machine-generated text detection using probability curvature,

    E. Mitchell, Y . Lee, A. Khazatsky, C. D. Manning, and C. Finn, “Detectgpt: Zero-shot machine-generated text detection using probability curvature,” arXiv preprint arXiv:2301.11305 , 2023

    2023 arXiv

  10. [17]

    Gltr: Statistical detection and visualization of generated text,

    S. Gehrmann, H. Strobelt, and A. M. Rush, “Gltr: Statistical detection and visualization of generated text,” arXiv preprint arXiv:1906.04043 , 2019

    1906 arXiv

  11. [18]

    Release strategies and the social impacts of language models,

    I. Solaiman, M. Brundage, J. Clark, A. Askell, A. Herbert-V oss, J. Wu, A. Radford, G. Krueger, J. W. Kim, S. Kreps, et al. , “Release strategies and the social impacts of language models,” arXiv preprint arXiv:1908.09203, 2019

    1908 arXiv

  12. [19]

    In- trinsic dimension estimation for robust detection of ai-generated texts,

    E. Tulchinskii, K. Kuznetsov, L. Kushnareva, D. Cherniavskii, S. Nikolenko, E. Burnaev, S. Barannikov, and I. Piontkovskaya, “In- trinsic dimension estimation for robust detection of ai-generated texts,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  13. [20]

    Dna-gpt: Divergent n-gram analysis for training-free detection of gpt- generated text,

    X. Yang, W. Cheng, Y . Wu, L. Petzold, W. Y . Wang, and H. Chen, “Dna-gpt: Divergent n-gram analysis for training-free detection of gpt- generated text,” arXiv preprint arXiv:2305.17359 , 2023

    2023 arXiv

  14. [21]

    Raidar: generative ai detection via rewriting,

    C. Mao, C. V ondrick, H. Wang, and J. Yang, “Raidar: generative ai detection via rewriting,” arXiv preprint arXiv:2401.12970 , 2024

    2024 arXiv

  15. [22]

    Can ai-generated text be reliably detected?,

    V . S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi, “Can ai-generated text be reliably detected?,” arXiv preprint arXiv:2303.11156, 2023

    2023 arXiv

  16. [23]

    Gptzero: An ai text detector,

    Tian, “Gptzero: An ai text detector,” 2023

    2023

  17. [25]

    Provable robust water- marking for ai-generated text,

    X. Zhao, P. Ananth, L. Li, and Y .-X. Wang, “Provable robust water- marking for ai-generated text,” arXiv preprint arXiv:2306.17439 , 2023

    2023 arXiv

  18. [26]

    Remark- llm: A robust and efficient watermarking framework for generative large language models,

    R. Zhang, S. S. Hussain, P. Neekhara, and F. Koushanfar, “Remark- llm: A robust and efficient watermarking framework for generative large language models,” arXiv preprint arXiv:2310.12362 , 2023

    2023 arXiv

  19. [27]

    Cater: Intellectual property protection on text generation apis via conditional watermarks,

    X. He, Q. Xu, Y . Zeng, L. Lyu, F. Wu, J. Li, and R. Jia, “Cater: Intellectual property protection on text generation apis via conditional watermarks,” Advances in Neural Information Processing Systems , vol. 35, pp. 5431–5445, 2022

    2022

  20. [28]

    Protecting intellectual prop- erty of language generation apis with lexical watermark,

    X. He, Q. Xu, L. Lyu, F. Wu, and C. Wang, “Protecting intellectual prop- erty of language generation apis with lexical watermark,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36, pp. 10758– 10766, 2022

    2022

  21. [29]

    Robust multi-bit natural language watermarking through invariant features,

    K. Yoo, W. Ahn, J. Jang, and N. Kwak, “Robust multi-bit natural language watermarking through invariant features,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 2092–2115, 2023

    2023

  22. [30]

    Deeptextmark: Deep learning based text watermarking for detection of large language model generated text,

    T. Munyer and X. Zhong, “Deeptextmark: Deep learning based text watermarking for detection of large language model generated text,” arXiv preprint arXiv:2305.05773 , 2023

    2023 arXiv

  23. [31]

    Watermarking text generated by black-box language models,

    X. Yang, K. Chen, W. Zhang, C. Liu, Y . Qi, J. Zhang, H. Fang, and N. Yu, “Watermarking text generated by black-box language models,” arXiv preprint arXiv:2305.08883 , 2023

    2023 arXiv

  24. [32]

    Watermarking of large language models,

    Aaronson, “Watermarking of large language models,” Aug. 2023

    2023

  25. [33]

    Generating steganographic text with lstms,

    T. Fang, M. Jaggi, and K. Argyraki, “Generating steganographic text with lstms,” arXiv preprint arXiv:1705.10742 , 2017

    2017 arXiv

  26. [34]

    Detecting fake content with relative entropy scoring.,

    T. Lavergne, T. Urvoy, and F. Yvon, “Detecting fake content with relative entropy scoring.,” Pan, vol. 8, no. 27-31, p. 4, 2008

    2008

  27. [35]

    Computer-generated text detection using machine learn- ing: A systematic review,

    D. Beresneva, “Computer-generated text detection using machine learn- ing: A systematic review,” inNatural Language Processing and Informa- tion Systems: 21st International Conference on Applications of Natural Language to Information Systems, NLDB 2016, Salford, UK, June 22-2...

    2016

  28. [36]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al., “Language models are unsupervised multitask learners,” OpenAI blog , vol. 1, no. 8, p. 9, 2019

    2019

  29. [37]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing , Association for Computational Linguistics, 11 2019

    2019

  30. [38]

    Paraphras- tic representations at scale,

    J. Wieting, K. Gimpel, G. Neubig, and T. Berg-Kirkpatrick, “Paraphras- tic representations at scale,” arXiv preprint arXiv:2104.15114 , 2021

    2021 arXiv

  31. [39]

    Sentence similarity learning by lex- ical decomposition and composition,

    Z. Wang, H. Mi, and A. Ittycheriah, “Sentence similarity learning by lex- ical decomposition and composition,” arXiv preprint arXiv:1602.07019, 2016

    2016 arXiv

  32. [40]

    Hcti at semeval-2017 task 1: Use convolutional neural network to evaluate semantic textual similarity,

    Y . Shao, “Hcti at semeval-2017 task 1: Use convolutional neural network to evaluate semantic textual similarity,” in Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017) , pp. 130–133, 2017

    2017

  33. [41]

    Sentence modeling via multiple word embeddings and multi-level comparison for semantic textual similarity,

    N. H. Tien, N. M. Le, Y . Tomohiro, and I. Tatsuya, “Sentence modeling via multiple word embeddings and multi-level comparison for semantic textual similarity,” Information Processing & Management , vol. 56, no. 6, p. 102090, 2019

    2019

  34. [42]

    Opt: Open pre-trained transformer language models,

    S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, C. Dewan, M. Diab, X. Li, X. V . Lin, et al. , “Opt: Open pre-trained transformer language models,” arXiv preprint arXiv:2205.01068 , 2022

    2022 arXiv

  35. [43]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  36. [44]

    Gpt-4o mini: advancing cost-efficient intelligence,

    OpenAI, “Gpt-4o mini: advancing cost-efficient intelligence,” July 2024

    2024

  37. [45]

    Billion-scale similarity search with gpus,

    J. Johnson, M. Douze, and H. J ´egou, “Billion-scale similarity search with gpus,” IEEE Transactions on Big Data , vol. 7, no. 3, pp. 535–547, 2019

    2019

  38. [46]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. , “Training language models to follow instructions with human feedback,”Advances in Neural Information Processing Systems , vol. 35, pp. 27730–27744, 2022

    2022

  39. [47]

    Maximum likelihood estimation of intrinsic dimension,

    E. Levina and P. Bickel, “Maximum likelihood estimation of intrinsic dimension,” Advances in neural information processing systems, vol. 17, 2004

    2004

  40. [48]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  41. [49]

    Gpt detectors are biased against non-native english writers,

    W. Liang, M. Yuksekgonul, Y . Mao, E. Wu, and J. Zou, “Gpt detectors are biased against non-native english writers,” Patterns, vol. 4, no. 7, 2023

    2023

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.